Written by Mostafa Elkourechy, Director of AI & Consulting Operations, SDG Group Middle East
THE DECISION LAYER Series - Article 2
Ask an AI assistant for last quarter’s revenue and you get a fluent, confident number in seconds. Ask a second tool and you may get a different one. The fix isn’t a better model. It is agreeing, once, what your numbers mean.
Picture a CFO who asks the company’s new AI assistant: “What was net revenue in the GCC last quarter?” The answer arrives in seconds, with a tidy chart and a confident tone. It does not match the number in the board pack.
Nobody made a mistake. The assistant counted revenue when invoices were issued. Finance counts it after returns and rebates. And “GCC” in the sales system includes a distributor-led market that finance reports separately. Each number is defensible. Together they are a problem, because the next question in the meeting is no longer about the business. It is about which number to believe.
When I look at an AI-assisted analytics project, this is the first thing I check, well before models or prompts come up. In the first article of this series, I argued that agents should draft and humans should release. The same rule applies to meaning. An AI can draft an answer, but a person must own the definition behind it.
Who owns the definition behind the answer?
Speed is cheap, Agreement is not.
The entrepreneur’s instinct is right: connect the assistant to the data and start asking. Value shows up in days, not quarters, and waiting for a perfect data model kills momentum.
The enterprise’s instinct is also right: a number that reaches a board, a bank or a regulator must mean one thing, and someone must be able to say what.
The evidence suggests the speed-first path hits a wall without the second. In 2023, researchers at data.world tested GPT-4 on business questions over a realistic enterprise SQL database. Asked directly, the model answered 16% correctly. When the same questions were asked over a knowledge graph that captured the business meaning of the data, accuracy rose to 54% (Sequeda, Allemang and Jacob, 2023). Models have improved since, so don’t read those percentages as today’s score. Read them as evidence of the mechanism: giving the model the meaning of the data changes how often it gets the answer right.
A newer benchmark points the same way. Spider 2.0, built from real enterprise database workflows, found that OpenAI’s o1-preview solved only 21.3% of its tasks, compared with 91.2% on the older and simpler Spider 1.0 (Lei et al., 2024).
Gartner reaches a similar conclusion from the executive side. It predicts that through 2026, organizations will abandon 60% of AI projects that are not supported by AI-ready data, and it found that 63% of organizations either lack the right data management practices for AI or are unsure whether they have them (Gartner, 26 February 2025).
In my view, the assistant isn’t slow or unintelligent. It is guessing which definition you meant.
The rule: Define once, answer everywhere
The design principle I work with is simple. Business meaning is written down once, owned by a named person, and every consumer, whether a dashboard, a report, a spreadsheet or an AI assistant, reads from that single definition.
1. Start with decisions, not data. List the ten or twenty decisions leadership makes repeatedly, then define only the metrics those decisions need. This is weeks of work, not a multi-year data program.
2. One metric, one definition, one owner. Each definition states the calculation, the filters, the time logic, the currency and the level of detail, in a form both people and machines can read. A named business owner, not IT, signs it off.
3. The assistant answers from definitions, not raw tables. It can choose among approved calculations. It can’t invent a new one on the fly.
4. No definition means no confident answer. If a question falls outside what has been defined, the right response is “I don’t have an approved definition for that,” followed by a proposal. The agent drafts the new definition and the owner releases it.
That last step is where the two layers of this series meet. The entrepreneur gets the speed of an assistant that answers immediately. The enterprise gets the control of numbers that always trace back to an owner.
Same word, different numbers
The pattern shows up in every function. Here are a few everyday terms that quietly hold more than one meaning:
|
Term |
Two defensible meanings |
What the shared definition settles |
|
Revenue |
Invoiced amount, or amount net of returns and rebates |
Which one the board sees, and when each is used |
|
Active customer |
Bought in the last 3 months, or in the last 12 |
The time window, and which channels count |
|
Margin |
Before, or after, logistics and promotion costs |
Which costs sit above the line |
|
Headcount |
Employees on payroll, or full-time equivalents including contractors |
Whether contractors and part-timers count |
|
Region |
Sales territory, or legal-entity reporting geography |
Which hierarchy rolls up to the board pack |
An AI assistant doesn’t resolve these ambiguities. It picks one, silently, and sounds just as sure either way.
“Won’t this slow us down?”
It’s the first objection I hear. The design is built to do the opposite, for three reasons.
- You start with the decisions that matter. Ten or twenty metrics, not the whole company.
- Every definition is written once and used everywhere. The same definition serves the dashboard, the monthly pack and the assistant, so the argument about what a number means is settled once, not repeated each month.
- Meetings get shorter. When a number arrives with its definition and its owner, the discussion moves from “whose number is right?” to “what do we do about it?”
Three questions to ask before you approve the next AI assistant
If you’re sponsoring or evaluating an AI assistant for analytics, put these to your team:
- Which definition produced that number, and who owns it? If nobody can say, you can’t defend the number in front of a board or an auditor.
- Do two people asking the same question in different words get the same answer? If not, the assistant is interpreting your business rather than reading it.
- What does the assistant do when it has no approved definition? If the answer is “it makes its best guess,” you have found the next board-pack argument.
The real edge is agreement, not speed
Fast answers are becoming a commodity. Every organization will have them. What will stay scarce is an assistant whose numbers the CFO, the auditor and the sales director all accept, because they agreed on the language first.
That agreement is part of what I call the decision layer. In my work at SDG Group, this is what I help clients do: agreeing the business definitions that matter, giving each a named owner, and making them the single source that dashboards and AI assistants alike answer from.
If your teams are getting fast answers and slow agreement, I’d like to compare notes. Reach out and let’s start this conversation.